Enhancing Road Safety with VideoMAE: A Self-Supervised Transformer Approach for Traffic Risk Detection

Road traffic accidents remain one of the leading causes of injury and death worldwide, placing a significant burden on public health systems and affecting the social and economic stability of many societies. The impact is particularly severe in developing regions, where limited infrastructure, insufficient traffic monitoring, and rapidly growing vehicle populations make effective road safety management more challenging. Many of these accidents occur in situations where potentially dangerous interactions between road users go unnoticed until it is too late, such as sudden braking, unsafe proximity between vehicles, or abrupt maneuvers in dense traffic conditions. In this paper, we investigate the use of a self-supervised deep learning approach to identify hazardous situations in traffic video data. Our method builds on the Video Masked Autoencoder (VideoMAE) framework, which is designed to learn meaningful spatiotemporal representations from video sequences. The model is first initialized using a VideoMAE architecture pretrained on a large-scale video dataset and subsequently fine-tuned using a smaller set of labeled traffic videos. This approach enables the model to capture meaningful motion and interaction patterns without requiring extensive frame-level annotations. Experimental evaluation demonstrates promising results, achieving 96% accuracy in identifying high-risk traffic situations. These findings indicate that self-supervised video representation learning offers a robust and scalable solution for traffic monitoring and accident prevention.